config_example.yaml

# ╔══════════════════════════════════════════════════════════════╗
# ║                   Camyla — Example Config                    ║
# ║                                                              ║
# ║  Structure:                                                  ║
# ║    1. top-level runtime  — data paths, execution env         ║
# ║    2. llm_endpoints      — LLM connection points (user-named)║
# ║    3. api_keys           — non-LLM API keys                  ║
# ║    4. llm_roles          — per-role overrides                ║
# ║    5. idea_generation    — Phase 1-3 idea generation         ║
# ║    6. experiment         — Stage 1-3 experiment execution    ║
# ║                                                              ║
# ║  Copy this file to `config.yaml` and fill in your own keys   ║
# ║  (or leave `api_key: ""` and export the matching env var).   ║
# ╚══════════════════════════════════════════════════════════════╝


# ================================================================
#  1. RUNTIME (top-level fields)
# ================================================================
data_dir: "/path/to/your/nnUNet_raw"
desc_file: null
copy_data: true
exp_name: run
generate_report: true
metric_tiebreak_threshold: 0.005

exec:
  timeout: 7200
  agent_file_name: runfile.py
  use_conda: true
  conda_env: "py310"
  format_tb_ipython: false

debug:
  stage3: false


# ================================================================
#  2. LLM ENDPOINTS
# ================================================================
llm_endpoints:
  my_openrouter:
    api_key: ""
    api_key_env: OPENROUTER_API_KEY
    base_url: "https://openrouter.ai/api/v1"
    model: "deepseek/deepseek-v3.2"
    temperature: 0.5

  my_dashscope:
    api_key: ""
    api_key_env: DASHSCOPE_API_KEY
    base_url: "https://coding.dashscope.aliyuncs.com/v1"
    model: "openai/glm-4.7"
    temperature: 0.9

  my_minimax:
    api_key: ""
    api_key_env: MINIMAX_API_KEY
    base_url: "https://api.minimaxi.com/v1"
    model: "openai/MiniMax-M2.5"
    temperature: 0.9


# ================================================================
#  3. API KEYS
# ================================================================
api_keys:
  s2:
    value: ""
    env: S2_API_KEY
  ncbi:
    value: ""
    env: NCBI_API_KEY


# ================================================================
#  4. LLM ROLES
# ================================================================
default_endpoint: my_openrouter

llm_roles:
  log_summary:        { temperature: 1.0 }
  journal:            { temperature: 1.0 }

  literature_backbone:       { model: google/gemini-3-flash-preview }
  core_theme:                { temperature: 1.0 }
  paper_extraction:          { temperature: 0.3 }
  citation_analysis:         { temperature: 0.3 }
  challenge_extraction:      { temperature: 0.3 }
  proposal_diagnostic:       { model: google/gemini-3-flash-preview, temperature: 0.3 }

  feedback:           { temperature: 0.9, max_tokens: 8192 }
  ablation_ideation:  { temperature: 0.9, max_tokens: 4096 }

  paper_agent:
    _default:            { temperature: 0.6 }
    PaperSummaryAgent:   { model: google/gemini-3-flash-preview, temperature: 0.3 }
    DatasetMetadataAgent: { model: google/gemini-3-flash-preview, temperature: 0.3 }
    IdeaRefineAgent:      { model: google/gemini-3-flash-preview, temperature: 0.5 }
    IdeaGenerationAgent:  { model: google/gemini-3-flash-preview, temperature: 0.8 }
    IdeaVerificationAgent: { model: google/gemini-3-flash-preview, temperature: 0.4 }
    ResultAnalysisAgent: { model: google/gemini-3-flash-preview, temperature: 0.5 }
    MockExperimentAgent: { model: google/gemini-3-flash-preview, temperature: 0.7 }
    MockAblationAgent:   { model: google/gemini-3-flash-preview, temperature: 0.7 }
    BibtexAgent:         { model: z-ai/glm-4.7,             temperature: 0.3 }
    MetadataAgent:       { temperature: 0.2 }
    OutlineAgent:        { temperature: 0.5 }
    ReviewAgent:         { temperature: 0.5 }

  paper_writing:
    plot_executor:   { endpoint: my_dashscope, temperature: 0.9 }
    latex_editor:    { endpoint: my_dashscope, temperature: 0.7 }
    writer:          { temperature: 0.6 }
    image_generator: { endpoint: my_openrouter, model: google/gemini-3.1-flash-image-preview, aspect_ratio: "16:9", image_size: "2K" }


# ================================================================
#  5. IDEA GENERATION
# ================================================================
idea_generation:
  enabled: true
  initial_queue_size: 1

  literature_search:
    target_papers: { phase1: 6, phase2: 6 }
    max_papers_per_search: 20
    max_iterations: 20
    min_year: { phase1: "2015-01-01", phase2: "2021-01-01" }
    sources: [semantic_scholar]
    filter_open_access: true
    enable_randomization: true

    challenge_discovery:
      enabled: true
      max_challenges: 3
      final_challenges: 3
      iterations: 6
      challenges_per_round: 3

  idea_generators:
    - name: creative_generator
      personality: "highly creative and novel approach focused"
      temperature: 1.0
      max_tokens: 4096
    - name: technical_generator
      personality: "technically rigorous and implementation focused"
      temperature: 1.0
      max_tokens: 4096
    - name: medical_generator
      personality: "medical imaging and healthcare AI focused"
      temperature: 1.0
      max_tokens: 4096

  assessment:
    temperature: 1.0
    max_tokens: 8192
    criteria:
      verifiability: { weight: 0.30, description: "Verifiability and experimental feasibility" }
      novelty:       { weight: 0.30, description: "Innovation and novelty level" }
      credibility:   { weight: 0.20, description: "Technical credibility and scientific reasonableness" }
      alignment:     { weight: 0.10, description: "Alignment with research objectives" }
      coherence:     { weight: 0.10, description: "Logical consistency and internal coherence" }
    scoring: { min: 0.0, max: 10.0, precision: 1 }

  research_proposal:
    num_proposals: 3
    modules_per_proposal: 3
    duplicate_check: true
    output_dir: "research_proposals"


# ================================================================
#  6. EXPERIMENT
# ================================================================
experiment:
  num_workers: 1
  limit_workers_to_gpus: true
  steps: 5

  stages:
    stage1_max_iters: 6
    stage2_max_iters: 3
    stage3_max_iters: 15           # Total step budget: ablation + debug + comparison variants

  stage2:
    max_iterations_per_innovation: 10

  stage3:
    max_ablations: 6
    prioritize_core_innovations: true

  code:
    candidates: [my_dashscope, my_minimax]
    max_tokens: 16384

  openhands:
    python_path: "/opt/conda/envs/py310/bin/python"
    pytest_path: "/opt/conda/envs/py310/bin/pytest"
    max_iterations: 100
    llm:
      timeout: 360
      max_output_tokens: 32768
      temperature: 0.9
    condenser:
      enabled: true
      max_size: 80
      keep_first: 2

  search:
    max_debug_depth: 3
    debug_prob: 0.5
    num_drafts: 1
    stage2_ucb:
      enabled: true
      c_puct: 1.5
      q_below_exponent: 0.5
      buggy_q_penalty: 0.2
      prior_power: 3

  proposal_refinement:
    enabled: true
    metric_threshold: 0.00
    max_refinements_per_substage: 1
    term_out_tail_chars: 1000